On 04.01.2012 13:14, Simon Riggs wrote:
> On Tue, Jan 3, 2012 at 11:28 PM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
>> Jim Nasby<jim@nasby.net> writes:
>>> On Jan 3, 2012, at 12:11 PM, Simon Riggs wrote:
>>>> This could well be related to the fact that DropRelFileNodeBuffers()
>>>> does a scan of shared_buffers, which is an O(N) approach no matter the
>>>> size of the index.
>>
>>> Couldn't we just leave the buffers alone? Once an index is dropped and that's pushed out through the catalog then
nothingshould be trying to access them and they'll eventually just get aged out.
>>
>> No, we can't, because if they're still dirty then the bgwriter would
>> first try to write them to the no-longer-existing storage file. It's
>> important that we kill the buffers immediately during relation drop.
>>
>> I'm still thinking that it might be sufficient to mark the buffers
>> invalid and let the clock sweep find them, thereby eliminating the need
>> for a freelist.
>
> My patch puts things on the freelist only when it is free to do so.
> Not having a freelist at all is probably a simpler way of avoiding the
> lock contention, so I'll happily back that suggestion instead. Patch
> attached, previous patch revoked.
So, you're proposing that we remove freelist altogether? Sounds
reasonable, but that needs to be performance tested somehow. I'm not
sure what exactly the test should look like, but it probably should
involve an OLTP workload, and large tables being created, populated to
fill the cache with pages from the table, and dropped, while the OLTP
workload is running. I'd imagine that the worst case scenario looks
something like that.
Removing the freelist should speed up the drop, but the question is how
costly are the extra clock sweeps that the backends have to do.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com
